Speech recognition system with generation of logarithmic values of feature parameters

Abstract

In a speech recognition system, logarithmic amplifiers avoid low level signal or noise problems by clamping all signals below an input threshold to a constant output level.

What is claimed is: 
     
       1. In a speech recognition system utilizing a generator of feature parameters, the improvement comprising: extractor means for producing a feature parameter signal from a signal containing speech information; and   converter means for transforming said feature parameter signal into a logarithmic signal wherein said converter means produces a first constant value when said feature parameter signal has a value less than a specific threshold value, and a second value corresponding to a logarithm of said feature parameter signal, when said feature parameter signal is at least as great as said specific threshold value.   
     
     
       2. In a speech recognition system, the improvement comprising: extractor means for producing a feature parameter signal from a signal containing speech information; and   converter means for transforming an input signal into a logarithmic signal, said input signal including said feature parameter signal, wherein said converter means produces said logarithmic signal in accordance with the following formula,   W=K when Pi<K       W=AlogPi+B; Pi≧K         where W equals the logarithmic signal, A and B are specified constants, Pi is said input signal and K is a specified threshold value.   
     
     
       3. The system of claim 2 wherein said input signal =Pi+N wherein Pi equals said feature parameter signal and N equals a noise signal. 
     
     
       4. A speech recognition system, comprising speech input means for detecting speech signals and producing a speech sector signal;   feature extractor means for producing a feature parameter signal from said speech sector signal;   logarithmic converter means for transforming said feature parameter signal into a logarithmic signal wherein said logarithmic converter means produces a first constant value as said logarithmic signal when said feature parameter signal has a value less than a specific threshold value, and produces a second value as said logarithmic signal, said second value being a logarithm of said feature parameter signal, when said feature parameter signal is at least as great as said specific threshold value; and   means for normalizing said logarithmic signal.   
     
     
       5. The system of claim 4 wherein said feature extractor means comprises: a plurality of bandpass filters each having different frequency bandwidths that receive said speech sector signal, each bandpass filter producing a filter output signal; a plurality of sample-hold circuits, one of said sample-hold circuits associated with each of said bandpass filters, each sample-hold circuit receiving a said filter output signal and producing a sample signal from said filter output;   analog switching means for sequentially accessing said sample signals from each of said sample-hold circuits and producing a composite analog signal; and   converter means for digitizing said composite analog signal, thereby producing said feature parameter signal.   
     
     
       6. The system of claim 4 wherein said speech input means, comprises: a microphone that detects speech signals;   an amplifier that produces amplified speech signals; and   a sector extractor that extracts speech sector signals from said amplified speech signals.   
     
     
       7. The system of claim 4 wherein said means for normalizing includes a power normalizer; and   a time axis normalizer.